Wildfire Prediction

This project uses A.I. models to predict whether a certain region is at risk of a wildfire. With a predictive tool like this, preventive measures can be taken to prevent fires.

Comment

Inspiration (Problem)

Wildfires are a big problem:

  • Harmful air pollutants such as “PM2.5, NO2, ozone, aromatic hydrocarbons, or lead” (WHO) which affect humans as well as wildlife
  • Releases a lot of Carbon Dioxide and other Greenhouse gasses contributing to climate change
  • Soil erosion that leads to damaged ecosystems
  • Contaminates water supplies and impacts aquatic ecosystems
  • Humans cause 85% of all wildfires yearly (National Park Service)
  • Over 7.5 million acres of the wild were consumed by flames in 2022

About our project (Solution)

We trained two AI models. Our 1st model takes in a 350x350 pixel satellite/topography image of a specific region and outputs whether the area is at risk of a wildfire. Our 2nd model takes in numerical weather parameters (temperature, humidity, and wind speed) and outputs whether the conditions signify a risk of a wildfire. To make interactions with the models more user-friendly, we integrated and established the models on a web application (localhost).

This project allows users to predict whether a certain region is at risk of being ravaged by a wildfire. With a predictive tool like this, preventive measures can be taken to prevent fires. An example of a preventive measure is controlled burning which is the practice of purposely setting fire to a bounded region to reduce the risk of rampant wildfires.

How we built it

Satellite/Topography Image Model:

  • Used Kaggle’s “Wildfire Prediction Dataset (Satellite Images)”
  • Images of the topography of regions in Quebec, Canada labeled either “wildfire” (where wildfires occurred “in the future”) or “nowildfire” (wildfires haven’t occurred)
  • https://www.kaggle.com/datasets/abdelghaniaaba/wildfire-prediction-dataset
  • Built a CNN model using Tensorflow and Keras API which classifies a given satellite/topography image into two classes which are “low risk of wildfires” (“nowildfires”) and “high risk of wildfires” (“wildfires”)
  • Trained and evaluated the model which has about 94% accuracy

Numerical Weather Parameters Model:

  • Used Kaggle’s dataset: https://www.kaggle.com/datasets/rtatman/188-million-us-wildfires which documents the location and time of various wildfires
  • Used Meteostat API to retrieve weather conditions (temperature, humidity, wind speeds) at the time and location of the wildfires (wildfire data)
  • Also generated some control data (no wildfire data) → random regions and their normal weather conditions
  • Built an AI model using Tensorflow and Keras API which outputs whether a weather condition (temp, humidity, wind speed) is of “low risk of wildfires” (“nowildfires”) and “high risk of wildfires” (“wildfires”)
  • Trained and evaluated the model

Used the Flask framework to create the web application that hosts the models and allows users to easily interact with the models.

Challenges we ran into

  • It was challenging to find the data we needed to train the AI models. For instance, the data the Satellite/Topography Image Model was trained on was only images of regions in Quebec, Canada.
  • The AI models were time-consuming to make since there were lots of parameters to tweak and neural network layers to order and rearrange to achieve the best results.
  • The weather APIs needed to access the data to train the Numerical Weather Parameters Model were difficult to work with and didn’t always have the data available
    • We ended up using the Meteostat API
  • Cleaning up and organizing large amounts of data and generating labels, etc. for the AI models (especially for the Numerical Weather Parameters Model) was extremely time-consuming and tedious
  • The Numerical Weather Parameters Model has about “100%” accuracy, though this is due to overfitting and insufficient data. The model actually isn’t very accurate as of now due to the challenges in finding proper data to train the model.

What we learned

  • Learned the complexity of making AI models
  • Learned that data and data collection is a huge part of making AI models
  • Learned how to use APIs
  • Learned how to implement Flask
  • Experience how it feels to collaborate as a team on a coding project

What's Next for Wildfire Prediction

  • Train the models on better and varied data
  • Have real-time satellite images and numerical weather data being passed into the models (though at a certain time interval so the models aren’t overloaded or running constantly)
  • In addition to classifying whether the real-time satellite images/data have a risk or not of wildfires, we would like for the output to trigger an alert and notify the relevant departments/organizations (fire departments, etc.) so action can be taken such as controlled burning, etc.

Built With

Share this project:

Updates